Datasets:

harvardMadsys
/

chutes_metadata

license: cc-by-4.0
pretty_name: A Year in LLM Serving — Chutes Production Trace
size_categories:
  - 1B<n<10B
tags:
  - llm-serving
  - inference
  - workload-trace
  - prefix-caching
  - load-balancing
configs:
  - config_name: trace
    default: true
    data_files:
      - split: train
        path: data/trace/train-*.parquet
  - config_name: chute_models
    data_files:
      - split: train
        path: data/chute_models.parquet

A Year in LLM Serving — Chutes Production Trace

This dataset contains one year of production LLM-serving traffic from Chutes, with 6,122,413,756 requests across 9,174 models. Each row represents one API invocation.

The dataset accompanies the paper A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing (arXiv:2608.13573).

The trace includes request timing, model and anonymized user identifiers, serving instances, token counts, time-to-first-token (TTFT), and prefix-cache reuse.

Subsets

Subset	Rows	Description
`trace` (default)	6,122,413,756	Request-level trace in 45 Parquet shards (~91 GB), ordered by `started_at`
`chute_models`	4,558	Model metadata, including model name, invocation count, and supported GPUs

Trace Schema

Field	Type	Description
`invocation_id`	int64	Request identifier
`function_name`	string	API endpoint
`chute_id`	string	Model identifier; join with `chute_models` for model metadata
`user_id`	string	Anonymized user identifier
`rehash_round`	int32	User anonymization rotation
`instance_id`	string	Serving instance identifier
`started_at`	timestamp[us]	Request start time
`completed_at`	timestamp[us]	Request completion time
`it`	float	Input tokens
`ot`	float	Output tokens
`ct`	float	Cached prefix tokens
`ttft`	float	Time to first token in seconds

Notes

Timestamps are normalized, preserving relative timing but not the original dates.
User identifiers are re-anonymized every three months. Use (user_id, rehash_round) as the user identifier.
ct is available for approximately the last 10 weeks of the trace and is null before then.
chute_models provides model names for 3,922 of the 9,174 chute_ids in the trace, covering 93.8% of all requests. The remaining chute_ids correspond to private or short-lived models.

Usage

DuckDB

Query the Parquet files directly from the Hugging Face Hub:

SELECT
    chute_id,
    count(*) AS requests,
    approx_quantile(ttft, 0.5) AS p50_ttft
FROM 'hf://datasets/harvardMadsys/chutes_metadata/data/trace/*.parquet'
GROUP BY chute_id
ORDER BY requests DESC
LIMIT 10;

Hugging Face Datasets

from datasets import load_dataset

trace = load_dataset(
    "harvardMadsys/chutes_metadata",
    "trace",
    split="train",
    streaming=True,
)

models = load_dataset(
    "harvardMadsys/chutes_metadata",
    "chute_models",
    split="train",
)

Download

hf download harvardMadsys/chutes_metadata \
    --repo-type dataset \
    --local-dir chutes_trace

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

If you use this dataset, please cite the paper below.

Citation

@article{nixon2026year,
  title={A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing},
  author={Nixon, William and Durbin, Jon and Standhartinger, Florian and Gunawi, Haryadi S and Yang, Juncheng},
  journal={arXiv preprint arXiv:2608.13573},
  year={2026}
}